See discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/265847509 An Efficient Point Cloud Management Method Based on a 3D R-Tree ARTICLE in PHOTOGRAMMETRIC ENGINEERING AND REMOTE SENSING · APRIL 2012 Impact Factor: 2.07 · DOI: 10.14358/PERS.78.4.373 CITATIONS 3 DOWNLOADS 52 VIEWS 54 5 AUTHORS, INCLUDING: Yeting Zhang Wuhan University 31 PUBLICATIONS 97 CITATIONS SEE PROFILE Xiao Xie Wuhan University 7 PUBLICATIONS 9 CITATIONS SEE PROFILE Available from: Xiao Xie Retrieved on: 25 July 2015
Special Issue: : Terrestrial Lidar/ASPRS Resource Directory
288 Apr i l 2012 Photogrammetric engineering & remote SenSing
Photogrammetric engineering & remote SenSing is the of-ficial journal of the American Society for Photogrammetry and Remote Sensing. It is devoted to the exchange of ideas and information about the applications of photogrammetry, remote sensing, and geographic information systems. The technical activities of the Society are conducted through the fol-lowing Technical Divisions: Geographic Information Systems, Photogram-metric Applications, Primary Data Acquisition, Professional Practice, and Remote Sensing Applications. Additional information on the functioning of the Technical Divisions and the Society can be found in the Yearbook issue of PE&RS. Correspondence relating to all business and editorial matters pertaining to this and other Society publications should be directed to the American Society for Photogrammetry and Remote Sensing, 5410 Grosvenor Lane, Suite 210, Bethesda, Maryland 20814-2144, including inquiries, mem-berships, subscriptions, changes in address, manuscripts for publication, advertising, back issues, and publications. The telephone number of the Society Headquarters is 301-493-0290; the fax number is 301-493-0208; email address is [email protected].
Pe&rS. PE&RS (ISSN0099-1112) is published monthly by the American Society for Photogrammetry and Remote Sensing, 5410 Grosvenor Lane, Suite 210, Bethesda, Maryland 20814-2144. Periodicals postage paid at Bethesda, Maryland and at additional mailing offices.
SubScriPtion. Effective January 1, 2012, the Subscription Rate for non-members per calendar year (companies, libraries) is $420 (USA); $436 for canada airmail (includes 5% for canada’s goods and Service tax (GST#135123065); $430 for all other foreign.
PoStmaSter. Send address changes to PE&RS, ASPRS Headquarters, 5410 Grosvenor Lane, Suite 210, Bethesda, Maryland 20814-2144. CDN CPM #(40020812)
memberShiP. Membership is open to any person actively engaged in the practice of photogrammetry, photointerpretation, remote sensing and geographic information systems; or who by means of education or profession is interested in the application or development of these arts and sciences. Membership is for one year, with renewal based on the anniversary date of the month joined. Membership Dues include a 12-month subscription to PE&RS valued at $68. Subscription is part of membership benefits and cannot be deducted from annual dues. Annual dues for Regular members (Active Member) is $135; for Student mem-bers it is $45 (E-Journal – No hard copy); for Associate Members it is $90 (see description on application in the back of this Journal). An additional postage surcharge is applied to all International memberships: Add $40 for canada airmail, and 5% for canada’s goods and Service tax (gSt #135123065); all other foreign add $60.00.
coPyright 2012. Copyright by the American Society for Photogrammetry and Remote Sensing. Reproduction of this issue or any part thereof (except short quotations for use in preparing technical and scientific papers) may be made only after obtaining the specific approval of the Managing Edi-tor. The Society is not responsible for any statements made or opinions expressed in technical papers, advertisements, or other portions of this publication. Printed in the United States of America.
PermiSSion to PhotocoPy. The appearance of the code at the bottom of the first page of an article in this journal indicates the copyright owner’s consent that copies of the article may be made for personal or internal use or for the personal or internal use of specific clients. This consent is given on the condition, however, that the copier pay the stated per copy fee of $3.00 through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, Massachusetts 01923, for copying beyond that permitted by Sections 107 or 108 of the U.S. Copyright Law. This consent does not extend to other kinds of copying, such as copying for general distribu-tion, for advertising or promotional purposes, for creating new collective works, or for resale.
Peer-Reviewed Articles
309 Laser Scanning in Heritage Documentation: The Scanning Pipeline and its ChallengesHeinz Rüther, Roshan Bhurtha, Christoph Held, Ralph Schröder, and Stephen WesselsThe aspects of the laser scanning pipeline and reports on advan-tages and challenges of terrestrial laser scanning for the docu-mentation of cultural heritage sites.
317 Cell-based Automatic Deformation Computation by Analyzing Terrestrial Lidar Point CloudsJing Wu, Pierre-Yves Gilliéron, and Bertrand MerminodA novel cell-based approach for deformation monitoring, includ-ing a hybrid model for deformation representation and an auto-matic procedure for deformation computation.
331 Automated Extraction of Road Markings from Mobile Lidar Point Clouds Bisheng Yang, Lina Fang, Qingquan Li, and Jonathan LiA promising solution for automatic extraction of road markings from mobile lidar point clouds.
339 Calibration and Kinematic Analysis of the Velodyne HDL-64E S2 Lidar SensorCraig GlennieThe suitability of the Velodyne HDL-64E S2 for high accuracy mo-bile scanning is presented by examining a kinematic calibration of the sensor.
349 Assessment of Available Rangeland Woody Plant Biomass with a Terrestrial Lidar SystemNian-Wei Ku, Sorin C. Popescu, R.J. Ansley, Humberto L. Perotto-Baldivieso, and Anthony M. FilippiDeveloping algorithms for estimating available woody plant bio-mass on rangelands by using a terrestrial lidar system.
363 Terrestrial Laser Scanning for Delineating In-stream Boulders and Quantifying Habitat Complexity MeasuresJonathan P. Resop, Jessica L. Kozarek, and W. Cully HessionAn algorithm developed for delineating in-stream boulders from terrestrial laser scanning data to provide automated measure-ments of structural complexity such as percent rock cover.
373 An Efficient Point Cloud Management Method Based on a 3D R-TreeJun Gong, Qing Zhu, Ruofei Zhong, Yeting Zhang, and Xiao XieAn efficient management and visualization approach based on 3DOR-Tree for very large point cloud data sets obtained from a vehicle-borne laser scanning system.
383 Tree Topology Representation from TLS Point Clouds Using Depth-First Search in Voxel SpaceAnita Schilling, Anja Schmidt, and Hans-Gerd MaasAssessment of forest relevant parameters based on a tree topol-ogy representation retrieved from TLS point clouds using depth-first search in voxel space.
Is your contact information current? Contact us at [email protected]
or log on to http://www.asprs.org/Member-Area/ to update your information. We value your membership.
xx
高亮
xx
附注
Jun GONG, Qing ZHU, Ruofei ZHONG, Yeting ZHANG, Xiao XIE, 2012. An Efficient Point Cloud Management Method Based on a 3D R-Tree. Photogrammetric Engineering & Remote Sensing, 78(4):373-381.
PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Ap r i l 2012 373
Jun Gong is with Department of Software, Jiangxi NormalUniversity, 99 ZiYang Road, Nanchang, 330022, P. R. China([email protected]).
Qing Zhu, Yeting Zhang, and Xiao Xie are with the StateKey Laboratory of Information Engineering in SurveyingMapping and Remote Sensing, Wuhan University, 129 LuoYu Road, Wuhan, 430079, P.R. China.
Ruofei Zhong is with College of Resources, Environment andTourism, Capital Normal University, 105 West Third RingRoad, Beijing, 100048, P.R. China.
Photogrammetric Engineering & Remote Sensing Vol. 78, No. 4, April 2012, pp. 373–381.
AbstractVehicle-borne laser-scanned point clouds have becomeincreasingly important 3D data sources in fields such asdigital city modeling and emergency response management.Aiming at reducing the technical bottlenecks of managementand visualization of very large point cloud data sets, thispaper proposes a new spatial organization method called3DOR-Tree, which integrates Octree and 3D R-Tree datastructures. This method utilizes Octree’s rapid convergenceto generate R-Tree leaf nodes, which are inserted directlyinto the R-Tree, thus avoiding time-consuming point-by-pointinsertion operations. Furthermore, this paper extends the R-Tree structure to support LOD (level of detail) models.Based on the extended structure, a practical data manage-ment method is presented. Finally, an adaptive controlmethod for LODS of point clouds is illustrated. Typicalexperimental results show that our method possesses quasi-real-time index construction speed, a good storageutilization rate, and efficient visualization performance.
IntroductionThe main requirements for 3D city models have been identi-fied as including high coverage, high fidelity, and being upto date, which are precisely the main shortcomings ofconventional 3D modeling approaches (Nebiker et al., 2010).Vehicle-borne laser scanning systems, such as Riegl’s VMX-250, Topcon’s IP-S2, and Optech’s Lynx, provide a goodsolution for the conflict between the rigorous requirements of3D modeling and finite human and financial resources. Suchmobile mapping systems combine multiple 3D laser scanners,GNSS, and IMU technologies and cameras to kinematicallyacquire 3D point clouds at centimeter-level density and 5 cmaccuracy (Barber et al., 2008; Leberl et al., 2010). Withaccurate calibration between the scanner and the camera,each laser point can be associated to a RGB value from theoptical image (Pu et al., 2009). Their raw data productionrates are on the order of 360 GB/hr, which means that thedata volume will approach the Terabyte level within threehours of operation. The unprecedented acquisition rate andhigh spatial resolution pose a considerable challenge for themanagement and visualization of the point clouds.
An Efficient Point Cloud Management Method Based on a 3D R-Tree
Jun Gong, Qing Zhu, Ruofei Zhong, Yeting Zhang, and Xiao Xie
As high-density 3D laser scanning becomes morepopular, there is an increasing demand for automaticpostprocessing of the large volumes of point cloud data thatare typically produced during a project. Most postprocess-ing, such as filtering, classification, and feature extraction,depends greatly on the performance of the organization andmanagement of the point clouds, which notably limitsfurther usages for Light Detection and Ranging (lidar) (Chenget al., 2011). The computer graphics community has devel-oped several specific data organization methods to acceleratevisualization efficiency and quality, but these methodsmostly focus on a single object and cannot tackle complexenvironments, such as those found in even small-scale laserscanning projects.
Because point cloud data are of large volume and high-resolution, adaptive visualization with level of detail (LOD)is a suitable strategy. Surfels and Qsplat are two promisingapproaches for LOD point rendering, but both require time-consuming preprocessing (Pfister et al., 2000; Rusinkiewiczand Levoy, 2000). Their unbalanced tree structures tend toinduce high tree depths and increasingly worse queryperformance. Improvements in the implementation of thesemethods have been obtained by adopting some advancedtechniques, such as parallel processing, GPU rendering, anddata caches (Wand et al., 2008).
Managing large-scale point clouds is a very activeresearch area in the spatial information community. Huang(2006) presents an organizational approach based on sequen-tial Quad-trees for airborne lidar data. A segment mappingtechnique is adopted to randomly extract LOD lidar data,which is linked with nodes at different levels. Although itimplements adaptive rendering for lidar data, randomextraction cannot ensure the best simplification result.Furthermore, the danger of producing an unbalanced indexstructure and its subsequent impact on point cloud manage-ment performance is also a factor. Lu et al. (2008) intro-duced a nested structure integrating Octree and binary treeto manage very large point cloud data sets. However, thespace partitioning is one-dimensional and cannot take intoconsideration the characteristics of 3D space. Kovac andZalik (2010) also adopt a hierarchical and out-of-coreapproach to manage point clouds. Ma and Wang (2011)distribute large point cloud data sets among multiple serversand boost up efficiency using parallel access.
3D R-Tree adaptively adjusts index structures according tothe real data. Therefore, object distribution has little influ-ence on R-Tree, making them a promising true 3D spatialaccess method (Zhu et al., 2007). However, when extendinginto 3D space, node overlapping easily induces multipath
queries, which becomes the main reason for low efficiency.Gong et al. (2011) introduced a global optimization mecha-nism and 3D cluster analysis to build the 3D R-Tree indexdynamically, a process that can solve the problem of seriousnode overlapping. This approach extends the R-Tree indexstructure and integrates LOD information into intermediatenodes. Theoretically, its dynamic updating and adaptiveadjusting capabilities are suitable for scattered and uneven 3Dpoint clouds. However, because of the algorithm’s complex-ity, so far, no related literature has been published.
This paper addresses that issue and is organized asfollows: following the Introduction, followed by a proposednew 3D spatial index method: 3DOR-Tree. The next Sectionintroduces how the 3DOR-Tree structure is extended tointegrate the LOD model, and next, a data organizationapproach based on 3DOR-Tree is presented for huge pointclouds. The next Section describes the adaptive control ofLODS for point cloud scenes, followed by the experimentalresults in comparison with other methods. Some conclu-sions are given in the final Section.
A New 3D Spatial Index Method: 3DOR-TreeIn vehicle-borne laser-scanning applications, the 3D pointcloud density is very high and can have a very unevenspatial distribution with large variations of height overrelatively short distances. These factors introduce majorchallenges to spatial index construction. In one cubic meter,there may exist any range between a few points and a fewhundreds of thousands of points. Traditional methods, suchas Cell and Quadtree, are based on 2D space and have beenextended into 3D space through 3D Cell and Octree. How-ever, with these structures, a great deal of empty nodes maybe produced because of the uneven spatial distribution ofthe data, which makes for very low spatial utilization andrapidly deteriorating query performance.
R-Tree construction methods can be divided intodynamic and static types. In static methods, bottom-upconstruction is adopted, which can ensure high efficiencyand spatial utilization, but it is difficult to generate theoptimal structure, and such methods do not support updat-ing of the tree. In dynamic methods, adaptive mechanismsare applied to ensure a reasonable tree shape, but theconstruction performance and spatial utilization are not asgood as than static methods. Although dynamic methods canbetter satisfy the spatial data management requirement,every point is inserted into an index structure using a seriesof complex operations, including node choosing and nodesplitting. Obviously, this is unreasonable for the very largedata sets commonly found in laser scanning projects. Thispaper adopts a new method, which integrates both staticand dynamic methods in order to achieve both high con-struction efficiency and adaptive updating.
The result is a hybrid spatial index method that inte-grates R-Tree and Octree structures (3D OCTR-Tree; abbreviatedas 3DOR-Tree). Before constructing a 3DOR-Tree structure,some parameters should be set by the user. The first are thefanout parameters of the 3D R-Tree (the maximum andminimum number of tuples in each R-Tree node, m and M).The second is the Octree’s convergence condition, which isthe maximum number of tuples in an Octree leaf node. Inour method, the Octree’s convergence condition equals themaximum fanout value of the 3D R-Tree.
The flow chart of the 3DOR-Tree construction algorithmis shown in Figure 1, and key issues of this algorithm arediscussed in detail below. In the process of 3DOR-Treeconstruction, an Octree is utilized to subdivide the 3D spacequickly, and subdivision continues until the point numberin the leaf nodes is less than or equal to the maximum
fanout value. If child nodes satisfy this fanout condition inthe R-Tree, their bounding box will be recalculated and theninserted into the 3D R-Tree as leaf nodes. If there exist childnodes whose point number is less than the minimum fanoutvalue, points in these child nodes are collected and sequen-tially reorganized into leaf nodes of the R-Tree and, finally,inserted into the 3DOR-Tree:
Algorithm Description: spatial index construction algorithm ofone point data set;Algorithm Input: R-Tree fanout parameters (m, M);Algorithm Output: 3DOR-Tree structure;
• Step 1: Find the minimal bounding box of all points (MinX,MinY, MinZ, MaxX, MaxY, MaxZ). From the beginning of(MinX, MinY, MinZ), find the minimal cubic bounding boxas Octree root node, Node. Create two new point arrays,named Array1 and Array2. Enter Step 2;
• Step 2: If point number in Node �= M, find its minimalbounding box and make it the root node of 3DOR-Tree, thenenter Step 9; If point number in Node � M, enter Step 3;
• Step 3: Subdivide Node into 8 child nodes Childi (I =0,1,…,7) evenly and distribute points in Node into childnodes, then enter Step 4;
• Step 4: Clear Array1. Traverse child nodes Childi (i =0,1,…,7). If point number in Childi � m, add its points intoArray1. Enter Step 5;
• Step 5: Let iPtNum = the point number in Array1. If iPtNum� 2*M, enter Step 6; If iPtNum is in (M, 2*M], the points inArray1 are evenly divided into two leaf nodes and insertedinto R-Tree, then enter Step 7; If iPtNum is in [m, M], thepoints in Array1 are reorganized into one leaf node andinserted into R-Tree, then enter Step 7; If iPtNum � m, insertthe points in Array1 into Array2, then enter Step 7;
374 Ap r i l 2012 PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING
Figure 1. Flow chart of the 3DOR-Tree constructionprocedure.
• Step 6: Let k = Rounding(iPtNum/iMax). The preceding (k-1)*M points are divided into (k-1) leaf nodes and insertedinto R-Tree. The number of the resting points, iRestNum,equals iPtNum-(k-1)*M. If iRestNum = M, the resting pointsare reorganized as one leaf node and inserted into R-Tree. IfiRestNum � M, divide the resting points evenly into two leafnodes and insert them into R-Tree. Enter Step 7;
• Step 7: Traverse every child node Childi (i =0,1,…,7). If thepoint number in Childi is in [m, M], find its minimalbounding box and insert it into R-Tree; If the point numberin Childi � M, let Node = Childi, then enter Step 3;
• Step 8: If all Octree subdivisions end, insert the points fromArray2 into R-Tree. Enter Step 9;
• Step 9: Exit.
The Octree is utilized to allocate neighboring points intothe same or adjacent nodes. A node is used as an insertionunit to insert points in bulk to avoid time-consumingincremental-insertion operations, and hence, significantlyboost index generation efficiency. The dynamic insertionmode allows the tree structure to have good spatial adapta-tion, which ensures a balanced tree structure and soundspatial utilization. Figure 2 shows the Octree structure of apoint cloud data set (splitting parameter is 100). Figure 3depicts the corresponding 3DOR-Tree structure (fanoutparameters are 40 and 100).
3DOR-Tree Extended Structure Concerning LoDsIn generic point cloud applications, the data volume is verylarge, with some projects containing a billion or more points.Such data sets usually far surpass the capabilities of theaverage computer system, especially if real-time interactionwith the data is desired. The most influential factor on systemperformance and data interaction is LOD; the greater the LOD tobe displayed, the greater the system performance must be.Because vehicle-borne laser scanning applications require highinteractivity, an efficient LOD strategy becomes a prerequisite,which means that the proper level of detail should be selectedto represent point cloud scenes in real time according to viewlength and software/hardware performance. Previous researchabout integrating R-Tree and LODS tried to utilize R-Tree’shierarchical structure to realize the dual functions of objectquery and LOD query (Kofler, 1998; Zlatanova, 2000). Theminimal bounding box of an R-Tree node is regarded as a lowLOD representation in those approaches, which obviouslycannot satisfy high-quality visualization.
A traditional R-Tree only manages object models in aleaf-node layer. The structure developed in this paperextends the management of object models to intermediatenodes, as is illustrated in Figure 4. A leaf-node layermanages all objects within the layer. A single representativeobject (for example, the one nearest to the centroid of thenode) is chosen and stored in the father node as the coarserobject model. With this strategy, the number of objects in afather node is equal to the number of child nodes.
PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Ap r i l 2012 375
Figure 2. Leaf node layer in Octree.
Figure 3. Leaf node layer in 3DOR-Tree.
Figure 4. 3DOR-Tree structure.
Based on the hierarchical structure of an R-Tree, the leafnodes denote the highest LOD, intermediate nodes denote amedium LOD, and the root node represents the lowest LOD.Every layer has one range, which includes the nearestdistance and the farthest distance. The ranges of neighboringlayers are seamless, and all nodes in the same layer have thesame range. When view distance is in the range of one node,point models in the node will be accessed and rendered. Inthe mode of panoramic representation, only points in theroot node are accessed. When the viewpoint is closer, theviewshed become smaller, and displayed details graduallyincrease. Figure 5 shows an example of the LOD representa-tion of a point cloud.
Point Cloud Organization Method Based on 3DOR-TreeBecause of practical constraints related to managing verylarge file sizes, large-scale point cloud projects are organizedin a project-cloud-point hierarchy. In the data-capturingprocess, vehicle-borne laser scanning systems typicallypartition the data into single point clouds of several millionpoints. The size of a point cloud file may amount to severalhundred MBytes. The collection of such point clouds in oneproject is called a “point cloud project.” Our experience hasshown that the point cloud project of one small village mayreach dozens of GBytes. A feature of the 3DOR-Tree structureis that it can manage such large point cloud projects throughimplementation in commercial DBMS.
In our file organization mode, one point cloud project isa file catalog that includes many point cloud files. Onepoint cloud file is a single binary-format file, whose formatis defined in Figure 6. One single point cloud file comprisesboth header and entity parts. The metadata of the pointcloud is stored in the header part, including version, datavolume, fanout parameters, the total number of points andlayers, centroid coordinates, compression flag, and the rootnode address. Point coordinates in the entity part aredouble-precision real values relative to the centroid coordi-nates, which allows them to be expressed as small valueswhile maintaining the coordinates’ precision. Moreover, ifthe range in each coordinate axis is less than 655.35 metersand centimeter-level precision is required, the coordinatescan be expressed as a 2-byte short integer after beingmultiplied by 100, which allows data to be compressed to25 percent of its original size.
A 3D R-Tree index structure is adopted to manage pointdata entities. To avoid repetitive storage, the representativepoints from child nodes are moved to father nodes so thatthe number of points in low-level nodes will be minus one.
To realize a cache mechanism, father nodes need torecord the storage address of child nodes, i.e., the offset ofchild nodes relative to the starting address of the pointcloud file. R-Tree storage can be classified as either breadthtraversal storage or depth traversal storage. The breadthtraversal storage sequence means that node data is stored insequence at each R-Tree level, beginning at the root layer.Nodes are recorded to the point file layer by layer. Theshortcoming of this method is that a father node cannot bedirectly stored with its child nodes. The depth traversalstorage sequence means that the root node is first recorded,followed by its subtrees, so that every node is recorded inthe same way as a root node. The problem with this struc-ture is that sibling nodes in middle layers cannot berecorded together. Figures 7 and 8 illustrate the principles ofbreadth traversal and depth traversal storage, respectively.
Only point data in the top layers is required to repre-sent the whole scene. When the viewpoint approaches localscene level, child nodes will be visited using their fathernode until the leaf nodes are reached. Therefore, neither ofthe two storage sequences can ensure that, on any occasion,points that are visited at the same time are stored together.To boost the efficiency, a hybrid scheme is adopted torecord the 3DOR-Tree.
Let the leaf layer be the 1st layer. When a project isopened, the whole scene is presented. If a panoramic viewrequires point data in the 3rd layer and above, points in theselayers should be recorded in breadth traversal sequence.Points in the lowest two layers are stored in depth traversalsequence. From a practical perspective, the total amount ofdata in the point cloud project determines the panoramicboundary layer. If the data volume is relatively small, thenthe layer can be lower, and vice versa. The principle of ourapproach is explained in Figure 9, in which the 3rd layer isthe boundary. Nodes that will probably be visited at thesame time are combined, thus benefiting from fast accesstime. Because father nodes record the addresses of all childnodes, a file mapping technique is easily adopted to accesspoints in any node from the root layer to the leaf layer,which is theoretically simple and practically efficient.
376 Ap r i l 2012 PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING
Figure 5. LOD representation of point clouds: (a) high LOD, (b) medium LOD, and (c) low LOD.
(a) (b) (c)
PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Ap r i l 2012 377
Figure 6. Data organization of massive point cloud data sets.
Figure 7. The principle of breadth traversal storage. Figure 8. The principle of depth traversal storage.
Figure 9. The principle of the hybrid storage approach.
Adaptive Visualization Method Based on 3DOR-TreeSoftware and hardware environments differ greatly, makingadaptive control methods helpful when dealing with a broadrange of computing systems. A critical aspect of datamanagement is how to manage such adaptive approaches.Usually, the importance of objects can be quantitativelydescribed by some rules, such as view length. Points arethen rendered based on importance until the particularsystem’s limit is reached.
LoD Definition ParametersIn the 3DOR-Tree model, LODS are distributed into the nodesin corresponding layers of the R-Tree. Controlling howvarious LODS are defined is performed using a set of defini-tion parameters, and they are here described in detail.
First, the functional range of every LOD must be properlydefined to ensure a stable quantity of visible points indifferent viewsheds. Take an ideal 3D scene, one in whichthe points are evenly distributed and the nodes in all R-Treelayers are also evenly distributed, as an example. Let theview line be straight down, and locate one critical view. Atthe view, LOD2 will be visible as rising up, so the visiblescene is fully represented in the highest LOD, LOD1. Thegreatest view length for the visible scene equals the maxi-mum value of LOD1’s functional range, d. In this way, thesecond critical view can be found at which the visible sceneis fully represented in LOD2. At this time, the farthest viewlength is approximately the farthest distance of LOD2, D.Figure 10 denotes the scope comparisons in the above twoviews. The fanout parameters in R-Tree, m and M, determinethat in a specific area, the ratio of node numbers in neigh-boring layers is 1: m ~ 1:M. In our method, the pointnumber in any node also fits the fanout parameters, whichmeans that the point number in every node is similar. Thearea ratio of the above two scenes is (D/d)2, and the ratio ofnode numbers in neighboring layers is 1:m ~ 1:M. To ensuresimilar node numbers in the two scenes, m � (D/d)2 � M.The farthest distance of adjacent LODS should meet thegeometric relationship. In the condition of m = 40 and M = 100, 6 � D/d � 10, which can ensure almost the samepoint numbers in different viewing fields.
In our method, three parameters are viewed as LODdefinition parameters, e.g., the number of R-Tree levels(LevelNum), the farthest distance of the highest LOD(FarDist), and the ratio of the farthest distances of adjacent
LODS (DistFactor). Suppose R-Tree has four levels, namely,LevelNum = 4, then let FarDist = 15 m and DistFactor = 8.The functional range of the 1st level is 0 ~ 15 m, the 2nd oneis 15 ~ 120 m, the 3rd one is 120 ~ 960 m, and the 4th one is960 ~ 8000 m (the farthest distance of the final level may beunlimited).
Adaptive Control of LoDs for a 3D SceneAccording to the preceding section, the range of every levelcan be adjusted by changing FarDist. When FarDist is madelarger, the range of every level also becomes larger. Hence,the displayed complexity of a 3D scene can be changed byadjusting FarDist. Adaptive control of the point cloud scenecan be achieved by adjusting FarDist. If the scene is to besimplified, FarDist needs to be less, and vice versa.
Quantitative control of the 3D scene utilizing LODparameters will be discussed below. When the viewpoint isclose to the ground and the view line is horizontal, theviewing field contains the most objects. Slod1 is the influen-tial area of LOD1, and Slod2 is that of LOD2, which are respec-tively calculated by Equations 1 and 2. Figure 11 illustratesthe influential area of LOD1 and LOD2:
(1)
(2)
where D is the farthest distance of the highest LOD (FarDist),K is the ratio of the farthest distances of neighboring LODS(DistFactor), and � is the horizontal angle of the viewfrustum.
Both Slod1 and Slod2 are in proportional relationship withD2. Suppose the 3D scene belongs to an ideal state in whichthe original point density is basically even and the pointdensity in every level is also even. In this condition, theoverlay area of every level decides the point number inside,so the processing cost is proportional to D2. In real-timeinteraction, D may be adjusted according to previous framesto deal with load variance. On the premise of a stable framerate, the richest LOD scene can be loaded and represented inthe limitation of the available main and video memory.
Experimental AnalysisThe performance of the 3DOR-Tree was tested on a sampledata set of a small village captured by one vehicle-bornelaser scanning system. The true color 3D data was acquiredin 20 minutes and included all roads, building façades,trees, and electricity lines. The data comprise 92 point
Slod2 � 1K 2 �12pD2*a/360,
Slod1 � pD2*a/360;
378 Ap r i l 2012 PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING
Figure 10. The scope comparison of LOD1 and LOD2.Figure 11. The comparison of the LOD1 andLOD2 influential areas.
clouds, with each cloud made up of approximately 2-millionpoints. The total data volume was more than 16 GB andapproximately 220 million points. The computing environ-ment was very modest: a laptop computer powered by anIntel Duo T7500 CPU and 1 GB of main memory.
Index Construction PerformanceThe first test was to assess the efficiency of indexing asingle point cloud of 2,426,454 points from the data set. Theperformance of the 3DOR-Tree structure was tested againstthat of a standard Octree and of a dynamic 3D R-Tree. Thesplit parameter in the Octree was set to 100, and fanoutparameters in the 3D R-Tree were set to 40 and 100. Thesesame parameters were set for the 3DOR-Tree. Two itemsrelating to construction efficiency were compared: the timefor construction and the depth of the resulting tree. Figures12 and 13 present the results.
In terms of speed, the Octree was constructed in justfive seconds, whereas the 3DOR-Tree took five times longer at25 seconds. Both these results could be considered toconstitute near-real-time performance. At 574 seconds, thedynamic 3D R-Tree took 23 times longer than the 3DOR-Treeand 115 times longer than the Octree. Inspection of theOctree’s structure showed an imbalance, with the depth ofleaf nodes ranging from 1 to 17. The depth of all leaf nodesin the 3DOR-Tree was 4, and the corresponding value for thedynamic 3D R-Tree was 5. It is easily understood from thetest results that the construction of the 3DOR-Tree satisfiesthe quasi-real-time requirement. Its tree depth is balancedand is less than the other two methods as, by its nature,almost all leaf nodes are full. Hence, the number of nodes isless so that the tree depth is smaller, which is of benefit tothe algorithm’s efficiency. It took approximately 40 minutesto construct the 3DOR-Tree index for the whole data set withthe Octree and dynamic 3D R-Tree being faster and slowerproportional to the single point cloud results.
Storage Space Utilization RateCompared with the CPU and the main memory, the speed ofexternal storage access is typically two orders of magnitudeslower, with the result that data volume affects data sched-uling performance and the user’s experience. As shown inFigure 14, when compared with a standard ASCII text file, asaving of over 90 percent can be achieved. When comparedwith a leading commercial point cloud software, Pointools™,a saving of 20 percent was observed.
Data Access and Visualization PerformanceWhen comparing the time it takes to open the file,Pointools™ took four minutes before the 220 million pointswere ready for user interaction.
In our method, with the fanout parameters of 40 and 100,point models in the 3rd level and higher do not exceed 105
points: (2*108/(40*40)). In the beginning of a project, onlypoints in the 3rd and higher levels are required. As theselevels are stored in breadth traversal sequence nodes, theselayers can be sequentially accessed, resulting in only a shortdelay before the project can be viewed. Experimental resultsshow that opening the project only took four seconds. Figure 15shows the 3rd level panoramic view of the sample project.
When stored in depth-first format, the project tookapproximately three minutes before a user was able tointeract with the point cloud.
The relatively low specification of the computingenvironment meant that the data cache should not exceed200 MB and that less than 5 million points be rendered in asingle frame. The data cache will be automatically clearedonce the 200 MB limit is exceeded. The farthest distance ofthe highest LOD was kept to approximately 15 meters, whichmeant the amount of data scheduled in a single frame wouldnot exceed 10 MB and would take less than two seconds torefresh. Because the rendering is part of a multithreadmechanism, the visualization task in the main thread willnot be influenced by such refresh times. Figure 16 is the LODdescription of the point cloud scene in which high detailsare represented in the near distance and few details arerepresented in the far distance.
Spatial Query Performance3D mapping and modeling is one of the key tasks to beformed on vehicle-borne laser scanning point clouds, and 3Dsnapping operations are crucial to an efficient workflow.Without an efficient spatial index, searching appropriatetarget points from potentially billions of candidates is a time-consuming process. When tested on a sample data set, it wasfound that the powerful indexing capacity of the 3D R-Treemakes snapping an instantaneous operation, which satisfiesuser requirements for interactive mapping. Other operations,such as moving through the point clouds and connectinglines, were also performed in real time. Figure 17 shows the
PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Ap r i l 2012 379
Figure 12. Comparison results of the spatial indexconstruction efficiency.
Figure 13. Comparison results of the tree depth. Figure 14. Comparison results of the spatial utilization.
380 Ap r i l 2012 PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING
Figure 15. Panoramic view of a point project.
Figure 16. LOD representation in the near distance.
Figure 17. 3D mapping operation based on the management platform.
effects of some simple mapping operations using thedeveloped platform.
ConclusionsThis paper introduces a new, fully 3D spatial index called3DOR-Tree and proposes an efficient management methodbased on LOD for very large point cloud data sets, such asthose acquired by vehicle-borne laser scanning technology.Based on a natural hierarchical structure, the LOD modelmanagement approach is useful for the adaptive visualiza-tion of massive point clouds.
With the rapid development of sensor technologies, asingle point created by a laser scanning system may possessfour or more properties; thus, the creation of semanticinformation by automatic classification and object extractionbrings new challenges to point cloud management. Seman-tics can be viewed as another type of LOD, and furtherresearch will aim at data fusion and the semantic manage-ment of point clouds, which will realize advanced integra-tion of panoramic images, vectors and semantics.
AcknowledgmentsThis study is supported by the National Basic ResearchProgram of China (973 Program, No. 2010CB731801 and2011CB302306) and the National Natural Science Founda-tion of China (41001222 and 41021061).
ReferencesBarber, D., J. Mills, and S. Smith-Voysey, 2008. Geometric valida-
tion of a ground based mobile laser scanning system, ISPRSJournal of Photogrammetry and Remote Sensing, 63(1):128–141.
Cheng, L., J. Gong, M. Li, and Y. Liu, 2011. 3D building modelreconstruction from multi-view aerial imagery and lidar Data,Photogrammetric Engineering & Remote Sensing, 77(2):125–139.
Gong, J., Q. Zhu, Y.T. Zhang, and X.M. Li, 2011. An efficient 3D R-tree extension method concerned with levels of detail, ActaGeodaetica et Cartographica Sinica, 40(2):249–255.
Huang, X.F., 2006. Research on 3D Building Model Extraction fromAirborne LIDAR Data, Ph.D. dissertation, Wuhan University,Wuhan, China, 156 p.
Kofler, M., 1998. R-trees for Visualizing and Organizing Large 3DGIS Databases, Ph.D. dissertation, Graz University of Technol-ogy, Austria, 156 p.
Kovac, B., and B. Zalik, 2010. Visualization of LIDAR datasets usingpoint-based rendering technique, Computers & Geosciences,36(11):1443–1450.
Leberl, F., A. Irschara, T. Pock, P. Meixner, M. Gruber, S. Scholz,and A. Wiechert, 2010. Point clouds: Lidar versus 3D vision,Photogrammetric Engineering & Remote Sensing, 76(10):1123–1134.
Lu, M.Y., and Y.J. He, 2008. Organization and indexing method for3D points cloud data, Geo-information Science, 10(2):190–194.
Ma, H., and Z. Wang, 2011. Distributed data organization andparallel data retrieval methods for huge scanner point clouds,Computer & Geoscience, 37(1):193–201.
Manolopoulos, Y., A. Nanopoulos, and A.N. Papadopoulos, 2005. R-Trees: Theory and Applications, Springer Press, Germany, 223 p.
Nebiker, S., S. Bleisch, and M. Christen, 2010. Rich point clouds invirtual globes - A new paradigm in city modeling?, Computer,Environment & Urban System, 34(6):508–517.
Pfister, H., M. Zwicker, J. van Baar, and M. Gross, 2000. Surfels: Surfaceelements as rendering primitives, Proceedings of ACM SIGGRAPH2000, 23–28 July, New Orleans, Louisiana, pp. 335–342.
Pu, S., and G. Vosselman, 2009. Knowledge based reconstruction ofbuilding models from terrestrial laser scanning data, ISPRSJournal of Photogrammetry & Remote Sensing, 64(6):575–584.
Rusinkiewicz, S., and M. Levoy, 2000. QSplat: A multiresolutionpoint rendering system for large meshes, Proceedings of ACMSIGGRAPH 2000, 23-28 July, New Orleans, Louisiana,pp. 343–352.
Wand, M., A. Berner, M. Bokeloh, and P. Jenke, 2008. Processingand interactive editing of huge point clouds from 3D scanners,Computer & Graphics, 32(2):204–220.
Zhu, Q., J. Gong, and Y.T. Zhang, 2007. An efficient 3D R-treespatial index method for virtual geographic environments, ISPRSJournal of Photogrammetry & Remote Sensing, 62(3):217–224.
Zlatanova, S., 2000. 3D GIS for Urban Development, Ph.D. dissertation,ITC, The Netherlands, 187 p.
PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Ap r i l 2012 381
